TOPOGRAPHIC FEATURE EXTRACTION FOR BENGALI ...Later, Bhattacharya et al. [7] [8] have also used the...

Signal & Image Processing : An International Journal (SIPIJ) Vol.2, No.2, June 2011

DOI : 10.5121/sipij.2011.2215 181

TOPOGRAPHIC FEATURE EXTRACTION

FOR

BENGALI AND HINDI CHARACTER IMAGES

Soumen Bag and Gaurav Harit

Department of Computer Science and Engineering, Indian Institute of Technology,

Kharagpur, West Bengal-721302, INDIA {soumen, gharit}@cse.iitkgp.ernet.in

ABSTRACT

Feature selection and extraction plays an important role in different classification based problems such as

face recognition, signature verification, optical character recognition (OCR) etc. The performance of OCR

highly depends on the proper selection and extraction of feature set. In this paper, we present novel

features based on the topography of a character as visible from different viewing directions on a 2D plane.

By topography of a character we mean the structural features of the strokes and their spatial relations. In

this work we develop topographic features of strokes visible with respect to views from different directions

(e.g. North, South, East, and West). We consider three types of topographic features: closed region,

convexity of strokes, and straight line strokes. These features are represented as a shapebased graph which

acts as an invariant feature set for discriminating very similar type characters efficiently. We have tested

the proposed method on printed and handwritten Bengali and Hindi character images. Initial results

demonstrate the efficacy of our approach.

KEYWORDS

Bengali, Character image, Convexity, Hindi, OCR, Thinning, Topographic feature, shape-based graph

1. INTRODUCTION

Feature selection and extraction has wide range of application for pattern recognition. It plays an

important role in different classification based problems such as face recognition, signature

verification, optical character recognition (OCR) etc. From past several decades, many feature

selection and extraction methods are reported [1] for Indian character recognition. The accuracy

rate of OCR depends on feature sets and classifiers [2] [3]. Next we begin a brief description

about few important feature sets used in optical character recognition for Bengali and Hindi

documents.

Chaudhuri and Pal [4] have proposed the first complete printed Bengali OCR in 1998. In this

method, nine different strokes are used as primary feature set for recognizing basic characters and

template-based features are used for recognizing compound characters. To recognize handwritten

basic and compound characters, Das et al. [5] have used shadow, longest run, and quad-tree based

features. Dutta and Chaudhury [6] have proposed topological features, such as junction points,

hole, stroke segments, curvature maxima, curvature minima, and inflexion points of character


182

images for performing printed and handwritten Bengali alpha-numeric character recognition.

Later, Bhattacharya et al. [7] [8] have also used the same topological feature set for handwritten

Bengali numerals. To detect convexity of Bengali numerals, Pal and Chaudhuri [9] have used

water-flow model. They also used topological and statistical features for preparing feature set.

Bhowmik et al. [10] have proposed a stroke-based feature set for recognizing Bengali

handwritten character images. In this method, ten stroke based features which indicate the shape,

size and position information of a digital curve with respect to the character image, are extracted

from character images to form the feature vector. Majumdar [11] has introduced a new feature

extraction method based on the curvelet transform of morphologically altered versions of an

original character image. Table 1 gives a summary of different feature sets used in Bengali OCR

systems.

Table 1. Different feature sets used in Bengali OCR systems

The first complete OCR on Hindi (Devanagari) is introduced by Pal and Chaudhuri [12]. In this

method, they have used structural and template features for recognizing basic, modified, and

compound characters. To recognize real-life printed documents of varying size and font, Bansal

and Sinha [13] have proposed statistical features. Later Pal et al. [14] and Kompalli-Setlur [15]

have used the gradient features for recognizing handwritten Hindi characters. Bajaj et al. [16]

have used density, moment of curve and descriptive component for recognizing Hindi

handwritten numerals. Sethi and Chatterjee [17] have proposed a set of primitives, such as, global

and local horizontal and vertical line segments, right and left slant, and loop for recognizing

handwritten Hindi characters. Sharma et al. [18] have used directional chain code information of

the contour points of the characters for recognizing handwritten Hindi characters. Table 2 gives a

summary of different feature sets used in Devanagari OCR systems.

Method Feature Set

Chaudhuri and pal [4] Structural and Template

Das et al. [5] Shadow, Longest run, and Quad-tree

Dutta and Chaudhury [6] Structural and Topological

Bhattacharya et al. [7] [8] Topological and Structural

Pal and Chaudhuri [9] Watershed, Topological, and Statistical

Bhowmik et al. [10] Stroke-based

Majumder [11] Curvlet coefficient


183

Table 2. Different feature sets used in Hindi OCR systems.

All of the above mentioned methods do not consider shape variation for extracting features. But

in Indian languages, a large number of similar shape type characters (basic and conjunct) are

present. From that point of view, we have proposed novel features based on the topography of a

character to improve the performance of existing OCR in Indian script, mainly for Bengali and

Hindi documents. The major features of the proposed method are listed as follows:

1. The main challenge to design an OCR for Indian script is to handle large scale shape

variation among different scripts. Strokes in characters can be decomposed into segments

which are straight lines, convexities or closed boundaries (hole). In our work we consider

the topography of character strokes from 4 viewing directions. In addition to the different

convex shapes formed by the character strokes, we also note the presence of closed

region boundaries.

2. The extracted features are represented by a shape-based graph where each node

contains the topographic feature, and they all are placed with respect to their centroids

and relative positions in the original character image.

3. The proposed methodology is applicable for different languages.

4. This approach is applicable for printed as well as handwritten text documents.

5. This topographic feature set helps to differentiate very similar type characters

(similarity in shape) in a proper way.

This paper is organized as follows. Section 2 describes the proposed topographic features

extraction method. Section 3 contains the experimental results. This paper concludes with some

remarks on the proposed method and future work in Section 4.

Method Feature Set Pal and Chaudhuri [12] Structural and Template

Bansal and Sinha [13] Statistical

Pal et al. [14] Gradient

Kompalli and Setlur [15] Gradient

Bajaj et al. [16] Density, Moment of curve, and Descriptive component

Sethi and Chatterjee [17] Watershed, Topological, and Statistical

Sharma [18] Directional chain code information of contour points


184

2. PROPOSED TOPOGRAPHIC FEATURE EXTRACTION

METHOD

In this section we describe our proposed feature extraction method based on measurement of

convexity of character strokes from different viewing directions. Figure 1 shows the system

architecture of the proposed method.

2.1. Preprocessing

Given a scanned document page we binarize it using Otsu's algorithm [19]. Currently we are

working with printed documents with all text content. For Bengali or Hindi documents the entire

word gets identified as a single connected component because of the matra/shiro-rekha (head

line) which connects the individual characters. For this case we separate out the individual

akshara within a word by using the character segmentation methods reported in [20] (for Bengali)

and [21] (for Hindi).

2.2. Thinning for Image Skeletonization

Before extracting the features, character images are converted to thinned (i.e., single pixel thick)

curve segments. But to retain proper shape of thinned character images (particularly for India

languages) is still a big challenge. Traditional thinning strategies lead to deformation of character

shapes, especially where the strokes get branched out; i.e. the junction points in the skeleton. Lot

of works have been done on thinning, but most of them are on English, Chinese, and Arabic

languages [22] [23] [24] [25] [26] [27]. From this point of view, Bag and Harit [28] have

proposed an improved medial-axis based thinning strategy for Indian character images. It uses

shape characteristics of text to determine the areas within the image region to stop thinning

partially. These regions are marked as junction points. This approach helps to preserve the local

features and true shape of the character. These resultant skeletons have stronger ability to cope

with all shortcomings of existing mask based thinning methods and would offer better models for

feature extraction and classification for OCR. Few results on printed Bengali and Hindi characters

are shown in Figure 2.

2.3. Extraction of Topographic Features

Topographic features are classified into three categories: closed region, convexity of strokes, and

straight line strokes. The convexity of curve is detected from different directions. Here we

consider four directions (North, South, East, and West) for convexity measurement.

Figure 1. System architecture of the proposed feature extraction method


185

Definition of convexity: The convexity of a shape is detected by checking its convergence

towards a single point or a cluster of points, connected by 4-connectivity. For any convexity

shape, if we move down word then we shall reach to a single point or a flat region which is a set

of 4-connected pixels. For detecting the stroke convexity, we have used this concept. Figure 3

shows the different convexities of character images detected from different view directions. In the

figure, different colors are used to mark convexities detected from different view directions (Red,

Blue, Green, and Magenta for North, South, East, and West respectively).

The following steps discuss the methodology to detect convexity from North direction.

1. Prepare a database containing different convex shape. Each convex shape is defined

according to their structural property. Figure 4 shows different shapes stored in the

database.

2. Detect horizontal-start (Hs), horizontal-end (He), vertical-start (Vs), and vertical-end

(Ve) points of the thinned image. Here [Hs, He] and [Vs, Ve] are the horizontal and vertical

limits of the input image respectively (Figure 5).

3. Given a thinned binary image, it is scanned from top to bottom and left to right, and

transitions from white (background) to black (foreground) are detected. The whole

scanning is done from Hs to He and Vs to Ve in horizontal and vertical directions

respectively.

Figure 2. Skeletonization results: (a) Input image; (b) Skeleton image.


186

Figure 3. (a) Skeleton image; (b) Different convex shapes marked by different colors (Red, Blue,

Green, and Magenta for North, South, East, and West respectively).

4. Suppose, a single scan from left to right generates a sequence of pixels {x1, x2, x3, ..., xi,

xi+1, ..., xn} where each pixel is a cut point between horizontal scan line and character

image. Now, if the two consecutive pixels (xi & xi+1) are not 4-connected neighbors, then

put them in an array of points P[1:N,1:2] where each cell P[i] contains one pixel with its

x, y coordinates in P[i][1] and P[i][2] respectively.

5. Continue the above steps for all remaining horizontal scan until we get a single point

or a set of horizontally 4-connected neighbor pixels with value ≥ ξ (set to 5).

6. The detected set of pixels is matched with the defined convex shapes stored in the

database to get a specific convex shape.

7. The above steps are repeated for detecting the convexity from remaining three

directions, i.e. South, East, and West. Only difference is that, for South direction, the scan

is done from bottom to top and left to right, and for East/West directions, scan is done

from top to bottom and right to left (for East)/left to right (for West) directions. Finally, a

set of different convex shapes are generated to prepare topographic feature set.

The above steps are used for detecting convexity of character strokes from different directions.

Now for detecting closed region, we use the concept of connected component analysis. If the

pixels of the stroke segment are connected (i.e. each pixel has two 8-connected neighbors), then

the stroke is identified as closed region.


187

Figure 4. Different convex shapes

Figure 5. Horizontal and vertical limits of an image

If the number of horizontally 8-connected neighbor pixels is ≥ € (set to 20), then we can say that

straight line is present. For Bengali and Hindi language, most of the characters have headline. By

observing the structural shape of these two types of characters, we can say that the straight line

will detect only from North direction. For this reason, we do not use straight detection method for

other three directions.

Figure 6 shows the different topographic features extracted from Bengali and Hindi thinned

character images.

2.4. Graphical Representation of Topographic Features

After detecting the convexity from four directions, we design a shape-based graph to represent

the feature set of a particular character. The steps are given below.

1. Suppose, a thinned character image has k number of topographic components {T1, T2,

..., Tk}.


188

Figure 6. Topographic features of thinned character images: (a) Skeleton image; (b)-(e) Features

from North, South, East, and West direction.

2. For each component, calculate the centroid (Xi, Yi) using the equation given below.

where Ni is the total number of black pixels of the ith topographic component, pij and p'ij

are the x-coordinate and y-coordinate of the ith pixel of the j

th topographic component

respectively.

3. Design an undirected graph G = (V, E), where V is the set of vertices containing

different topographic features and E is the set of edges. All vertices are placed based on

the coordinates of there centroid. Now add edges among these vertices with respect to

their relative positions in the thinned character image.


189

Figure 7 shows the shape-based graph of thinned Bengali and Hindi character images. This

graphical model gives a clear pictorial difference among different feature set of different

character images.

Figure 7. (a) Skeleton image; (b) Shape-based graph.

3. EXPERIMENTAL RESULTS We collected characters from several heterogeneous printed and handwritten documents of

Bengali and Hindi. The number of characters in the testing set is 2550 (1300 for Bengali and

1250 for Hindi). All the characters are collected in a systematic manner from printed and

handwritten pages scanned on a HP scanjet 5590 scanner at 300 dpi. At first, images were thinned

by Bag and Harit medial-axis based thinning method [28]. This thinning method has the

capability to preserve the shape of character images at junction points and end points. Then we

tested our proposed feature extraction method on these thinned images to get topographic


190

features. Figure 8 and 9 show the shape-graph representation of topographic features of printed

Bengali and Hindi character image. The algorithm is implemented in C++ programming language

using OpenCV 2.0 on Unix/Linux platform.

We applied this proposed method on handwritten Bengali and Hindi character images. Figure 10

shows few results. It is observed that the performance is satisfactory for handwritten characters

also. The proper recognition of handwritten character is still a big challenge.

This new feature set can discriminate characters those are very similar in shape. Figure 11 shows

few Hindi characters very similar in shape. After extracting topographic features and generating

shape-based graph using them, it is shown that these characters are distinguishable with respect to

their topographic feature set. Form this point of view; we can conclude that this new approach

will help to improve the accuracy rate of existing Bengali and Hindi OCR systems.

In future, we shall use these topographic features for optical character recognition for handwritten

and printed. The brief outline of the recognition method is given here. To identify different

topographic features, we give a unique number to each shape. Suppose, S00 is marked as 1, S01 is

as 2, and so on. Now we build a one dimensional feature set, containing different shape-ids

corresponding to each shape of a particular character. Repeating the same procedure, we prepare

a training feature set containing topographic features of different character images. Now for

recognition, we apply the same procedure on test sample to extract the features and match the

shape-ids for checking its presence in the training database.

4. CONCLUSION In this paper, we have proposed a novel topographic feature extraction method for Indian OCR

systems. This feature set captures close region, convexity of stroke from different direction, and

flat region of thinned character images. The extracted features are represented as a shape-based

graph and acts as an invariant feature set for discriminating similar type character images. The

proposed method is tested on printed and handwritten Bengali and Hindi characters and we have

obtained promising results. In future, we shall extend our work to extract topographic features for

other popular Indian languages and make it as a script independent feature set for designing

multi-lingual OCR in Indian scripts.


191

Figure 8. Topographic features of printed Bengali character images: (a) Input image; (b)Skeleton

image; (c) Shape-based graph.


192

Figure 9. Topographic features of printed Hindi character images: (a) Input image; (b) Skeleton

image; (c) Shape-based graph


193

Figure 10. Topographic feature extraction of handwritten character images: (a) Input image; (b)

Skeleton image; (b) Shape-based graph.


194

Figure 11. Discrimination between Hindi characters (similar in shape) using shape-based graphs

of their topographic features.

REFERENCES

[1] Trier, O. D., Jain, A. K., & Taxt, T., (1996) “Feature extraction methods for character recognition-A

survey”, Pattern Recognition, Vol. 29, No. 4, pp 641-662.

[2] Pal, U., Wakabayashi, T., Kimura, F., (2009) “Comparative study of Devnagari handwritten character

recognition using different feature and classifiers”, International Conference on Document Analysis and

Recognition, pp 1111-1115.

[3] Dalal, S., & Malik, L., (2008) “A survey of methods and strategies for feature extraction in handwritten

script identification”, International Conference on Emerging Trends in Engineering and Technology, pp

1164-1169.

[4] Chaudhuri, B. B. & Pal, U., (1998) “A complete printed Bangla OCR system”, Pattern Recognition,

Vol. 31, No. 5, pp 531-549.

[5] Das, N., Das, B., Sarkar, R., Basu, S., Kundu, M., & Nasipuri, M., (2010) “Handwritten Bangla basic

and compound character recognition using MLP and SVM classifier”, Journal of Computing, Vol. 2, No. 2,

pp 109-115.


195

[6] Dutta, A. & Chaudhury, S., (1993) “Bengali alpha-numeric character recognition using curvature

features”, Pattern Recognition, Vol. 26, No. 12, pp 1757-1770.

[7] Bhattacharya, U., Das, T. K., Datta, A., Parui, S. K., & Chaudhuri, B. B., (2002) “A hybrid scheme for

handprinted numeral recognition based on a self-organizing network and MLP classifiers”, International

Journal of Pattern Recognition and Artificial Intelligence, Vol. 16, No. 7, pp 845-864.

[8] Bhattacharya, U., Das, T. K., Datta, A., Parui, S. K., & Chaudhuri, B. B., (2002) “Recognition of

handprinted Bangla numerals using neural network models”, AFSS International Conference on Fuzzy

Systems, pp 228-235.

[9] Pal, U. & Chaudhuri, B.B., (2000) “Automatic recognition of unconstrained off-line Bangla handwritten

numerals”, International Conference on Advances in Multimodal Interfaces, pp. 371-378.

[10] Bhowmik, T. K., Bhattacharya, U., & Parui, S. K., (2004) “Recognition of Bangla handwritten

characters using an MLP classifier based on stroke features”, International Conference on Neural

Information Processing, pp 814-819.

[11] Majumdar, A., (2007) “Bangla basic character recognition using digital curvelet transform”, Journal of

Pattern Recognition Research, Vol. 2, No. 1, pp 17-26.

[12] Pal, U. & Chaudhuri, B. B., (1997) “Printed Devnagari script OCR system”, Vivek, Vol. 10, pp 12-24.

[13] Bansal, V. & Sinha, R. M. K., (2000) “Integrating knowledge sources in Devanagari text recognition

system”, IEEE Transactions on Systems, Man, and Cybernetics—Part A: Systems and Humans, Vol. 30,

No. 4, pp 500-505.

[14] Pal, U., Sharma, N., Wakabayashi, T., & Kimura, F., (2007) “Offline handwritten character

recognition of Devnagari script”, International Conference on Document Analysis and Recognition, pp

496-500.

[15] Kompalli, S. & Setlur, S., (2006) “Design and comparison of segmentation driven and recognition

driven Devanagari OCR”, International Conference on Document Image Analysis for Libraries, pp 96-102.

[16] Bajaj, R., Dey, L., & Chaudhury, S., (2002) “Devnagari numeral recognition by combining decision of

multiple connectionist classifiers”, Sadhana, Vol. 27, pp 59-72.

[17] Sethi, K. & Chatterjee, B., (1977) “Machine recognition of constrained hand-printed Devnagari”,

Pattern Recognition, Vol. 9, pp 69-77.

[18] Sharma, N., Pal, U., Kimura, F., & Pal, S., (2006) “Recognition of offline handwritten Devnagari

characters using quadratic classifier”, Indian Conference on Computer Vision, Graphics and Image

Processing, pp 805-816.

[19] Gonzalez, R. C. & Woods, R. E., (2008) Digital Image Processing, Prentice Hall, USA.

[20] Pal, U. & Datta, S., (2003) “Segmentation of Bangla unconstrained handwritten text”, International

Conference on document Analysis and Recognition, pp 1128-1132.

[21] Ma, H. & Doermann, D., (2003) “Adaptive Hindi OCR using generalized hausdorff image

comparison”, ACM Transactions on Asian Language Information Processing, Vol. 2, pp 193- 218.

[22] Lam, L., Lee, S.W., & Suen, C.Y., (1992) “Thinning methodologies-A comprehensive survey”, IEEE

Transactions on Pattern Analysis and Machine Intelligence, Vol. 14, No. 9, pp 869-885.


196

[23] Huang, L., Wan, G., & Liu, C., (2003) “An improved parallel thinning algorithm”, International

Conference on Document Analysis and Recognition, pp 780-783.

[24] Datta, A. & Parui, S. K., (1994) “A robust parallel thinning algorithm for binary images”, Pattern

Recognition, Vol. 27, No. 9, pp 1181-1192.

[25] Melhi, M., Ipson, S. S., & Booth, W., (2001) “A novel triangulation procedure for thinning hand-

written text”, Pattern Recognition Letters, Vol. 22, pp 1059-1071.

[26] Vincze, M. & Kovari, B., (2009) “Comparative survey of thinning algorithms”, International

Symposium of Hungarian Researchers on Computational Intelligence and Informatics, pp 173-184.

[27] Zhu, X. & Zhang, S., (2008) “A shape-adaptive thinning method for binary images”, International

Conference on Cyberworlds, pp 721-724.

[28] Bag, S., Harit, G., (2010) “A medial axis based thinning strategy and structural feature extraction of

character images”, IEEE International Conference on Image Processing, pp 2173-2176.

Date post:	14-Oct-2020
Category:	Documents
Upload:	others
View:	1 times
Download:	0 times

TOPOGRAPHIC FEATURE EXTRACTION FOR BENGALI ...Later, Bhattacharya et al. [7] [8] have also used the...

Documents